learninghouse 0.10.1

learningHouse - Teach your smart home everything

learningHouse Service

Introduction

learningHouse Service provides machine learning algorithms based on scikit-learn python library as a RESTful API, with the purpose to give smart home fans an easy possibility to teach their homes.

Add the moment this project is in a very early state. Please share your ideas what you want to teach your home by opening an issue. Really looking forward for your feedback.

Installation

Install and update using pip.

pip install -U learninghouse

Install and update using docker

docker pull learninghouseservice/learninghouse:latest

Prepare configuration directory

mkdir -p models/config
mkdir -p models/training
mkdir -p models/compiled

The config directory holds the model configuration as json-file.

The directories training and compiled are used by service to store data.

Training data is stored as csv file, trained models are stored as object dump.

Configuration of models

Configuration is stored in json format.

General configuration

In general send data of all sensors to learningHouse Service especially when training your models. The service will save all data fields even if they are not used in current model configuration as a feature. This will give you the possibility to choose different features later on to improve your model after some training (There will be some service later on to support you with this improvment).

In general there are two different data types your sensor data can be divided in. Numerical data can be processed directly by your models. Categorical data has to be preproccesed by the service to be used as a feature. Categorical data can be identified by a simple rule:

Non numerical values or if you can give each numerical value a term to describe it.

Some examples for categorical data:

• pressure_trend: Values of 'falling', 'rising', 'consistent'
• month_of_year: 1 ('January'), 2 ('Februrary'), ...

To use the data of your sensors as features in your models you have to give the service information about the data type. For this put a sensors.json to the directory model/config. List all your sensors and their data type.

Example content of sensors.json:

{
    "azimuth": "numerical",
    "elevation": "numerical",
    "rain_gauge": "numerical",
    "pressure": "numerical",
    "pressure_trend_1h": "categorical",
    "temperature_outside": "numerical",
    "temperature_trend_1h": "categorical",
    "light_state": "categorical"
}

Example model

The model decides whether it is so dark that the light has to be switched on. It uses the sun azimuth and sun elevation, the rain gauge and the one hour trend of air pressure. The data of the other senors (pressure, temperature_outside, light_state) isn't used in this example. It use a machine learning algorithm called RandomForestClassifier.

Store a darkness.json in models/config directory with following content:

{
    "estimator": {
        "typed": "classifier",
        "estimators": 100, 
        "max_depth": 5
    },
    "features": ["azimuth", "elevation", "rain_gauge", "pressure_trend_1h_falling"],
    "dependent": "darkness",
    "dependent_encode": true,
    "test_size": 0.2
}

Configuration parameter

Estimator

LearningHouse Service can predict values using an estimator. An estimator can be of type classifier which fits best for your needs if you have somekind of categorical output like in the example true and false. If you want to predict a numerical value for example the setpoint of an heating equipment use the type regressor instead.

For both types learningHouse Service uses a machine learning algorithm called random forest estimation. This algorithm builds a "forest" of decision trees with your features and takes the mean of the prediction of all of them to give you a best result. For more details see the API description of scikit-learn:

Estimator type	API Reference
RandomForestRegressor	https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.html#sklearn.ensemble.RandomForestRegressor
RandomForestClassifier	https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html#sklearn.ensemble.RandomForestClassifier

You can adjust the amount of decision trees by using estimators (default: 100) option. And the maximum depth of each tree by using max_depth (default: 5) option. Both options are optional. Try to resize this value to optimize the accuracy of your model.

Features

The list of features is required and holds the names of the sensor data the model uses to take a decision. Categorical data as mentioned above will be preprocessed by the service and is divided to one column by each known value. You can use each column as a seperate feature. There will be a service later on to help you to choose the set of best features. Meanwhile use the ones you think they maybe have influence on the decision as a first try and start training your house with all sensor data as mentioned above.

Dependent variable

The dependent variable is the one that have to be in the training data and which is predicted by the trained model.

The dependent variable has to be a number. If it is not a number, but a string or boolean (true/false) like in the example. For this set dependent_encode to true.

Test size

LearningHouse service only uses a part of your training data to train the model. The other part specified by test_size will be used to score the accuracy of your model.

Give a percentage by using floating point numbers between 0.01 and 0.99 or a absolute number of data points by using integer numbers.

For the beginning a test_size of 20 % (0.2) like the example should be fine.

The accuracy between 80 % and 90 % between is a good score to gain. Below your model is kind of underfitted and above overfitted, which both make it not working well for new data points to be predicted. You can try to change the set of used features to gain a better accuracy.

Training of the model will start, when there are at least 10 data points.

Run service

In the console type learninghouse to start the service in development mode. By default the service will run on port 5000. In development mode you will see some log information.

To start service in production mode and specify listen address and port use following command:

learninghouse --production --host 127.0.0.1 --port 5001

For more or less log output set the verbosity level of the service to DEBUG, INFO (default), WARNING, ERROR or CRITICAL

learninghouse --production --host 127.0.0.1 --port 5001 --verbosity DEBUG

Run with docker:

docker run --name learninghouse --rm -v models:/learninghouse/models -p 5000:5000 learninghouseservice/learninghouse:latest

Like above you can adjust the VERBOSITY_LEVEL by adding it as an environment variable.

Train model

For training send a PUT request to the service:

# URL is http://host:5000/training/:modelname
curl --location --request PUT 'http://localhost:5000/training/darkness' \
--header 'Content-Type: application/json' \
--data-raw '{
    "azimuth": 321.4441223144531,
    "elevation": -19.691608428955078,
    "rain_gauge": 0.0,
    "pressure": 971.0,
    "pressure_trend_1h": "falling",
    "temperature_outside": 23.0,
    "temperature_trend_1h": "rising",
    "light_state": false,
    "darkness": true
}'

You can send either a field timestamp with your dataset containing a UNIX-Timestamp or the service will add this information with its current time. The service generate some further time relevant fields inside the training dataset you can although use as features. These are month_of_year, day_of_month, day_of_week, hour_of_day and minute_of_hour

If one of your sensors is not working at the moment and for this not sending a value the service will add a value by using the following rules. For categorical data all categorical columns will be set to zero. For numerical data the mean of all known training set values (see Test size) for this feature will be assumed.

As mentioned above there will be a service feature which helps you to choose the features used in your model from time to time to improve the model. For this always send the data of every sensor when train your model. The service will store this values for possible future use.

To train the model with existing data for example after a service update use a POST request without data:

# URL is http://host:5000/training/:modelname
curl --location --request POST 'http://localhost:5000/training/darkness'

To get the information about a trained model use a GET request:

# URL is http://host:5000/info/:modelname
curl --location --request GET 'http://localhost:5000/info/darkness'

Prediction

To predict a new data set with your model send a POST request:

# URL is http://host:5000/info/:modelname
curl --location --request POST 'http://localhost:5000/prediction/darkness' \
--header 'Content-Type: application/json' \
--data-raw '{    
    "azimuth": 321.4441223144531,
    "elevation": -19.691608428955078,
    "rain_gauge": 0.0,
    "pressure_trend_1h": "falling"
}'

If one of your sensors used as feature in the model is not working at the moment and for this not sending a value the service will add this by using following rules. For categorical data all categorical columns will be set to zero. For numerical data the mean of all known training set values (see Test size) for this feature will be assumed.